REFINE: LLM Refinement over Budgeted Text-Attributed Graphs for Personalized Medical Concept Representation

To address the issue of existing encoders ignoring patients’ specific clinical backgrounds, the study proposes a framework called REFINE, which aims to optimize large language models using budgeted text attribution graphs to achieve personalized medical concept representation. This framework constructs patient-specific time-series graphs based on a global text attribution knowledge graph. Through sequential reinforcement learning strategies, it selects personalized graph expansion budgets for each observed code. Subsequently, a heterogeneous graph neural network captures relational and structural dependencies, and the frozen large language model uses graph-aware soft prompts for semantic refinement. Experiments on the MIMIC-III and MIMIC-IV datasets show that REFINE can continuously improve the performance of various electronic medical record backbone models, outperforming strong baseline methods, and demonstrating robust improvements in scenarios with component ablation, graph selection, and data insufficiency.